HyperAgent

HyperAgent is a sandboxed code-acting AI agent runtime: it writes JavaScript handlers, validates them, and runs them inside hardware-isolated Hyperlight micro virtual machines using the GitHub Copilot SDK.

It is built for useful, bounded work: data analysis, document generation, API workflows, secure file output, and tool use through plugins and MCP servers. The model can write code, but the code runs in a sandbox with no direct filesystem, shell, or network access unless you explicitly enable narrowly scoped host capabilities.

# Authenticate with GitHub; Copilot access is required.
gh auth login

# Install and run. Requires Node.js 22+.
npm install -g @hyperlight-dev/hyperagent
hyperagent --skill pptx-expert --profile web-research \
  --prompt "Create a presentation on the NASA Artemis II mission \
  include lots of statistics and data, use an appropriate theme \
  and color scheme for the subject, make it stunning"

What that gets you:

Instead of	HyperAgent gives you
Shell-first automation	Code-first handlers validated and run in a micro-VM
Ambient filesystem access	Path-jailed read/write plugins
Ambient network access	Domain-scoped fetch with SSRF checks
Ad hoc tool calls	Normal JavaScript APIs for approved capabilities
One-off generated scripts	Reusable handlers and modules
Hidden agent state	Explicit shared state, transcript logs, debug logs, and timing logs
Trust-me execution	Tool gating, code validation, plugin approval, and MCP review

HyperAgent is designed for useful, bounded jobs such as data pipelines, API exploration, reports, PowerPoint decks, PDFs, Excel workbooks, and workflows that connect approved MCP servers. Add --auto-approve only for trusted non-interactive runs; otherwise HyperAgent can pause for approval before enabling plugins, connecting MCP servers, or taking write-capable actions.

GitHub